Research

My research focuses on visual localization in GPS-denied environments, combining deep learning with geometric reasoning to build accurate and deployable positioning systems using only passive sensors.

LanBLoc: Landmark-Based Visual Localization

A passive visual localization framework that replaces traditional radio-based anchors with pre-mapped visual landmarks.

• Landmark detection using YOLO-based deep learning models
• Stereo vision for metric distance estimation
• Trilateration for position computation from multiple landmark observations
• Geometric optimization for refined, accurate estimates

Achieves significantly lower localization error compared to SLAM/VO baselines and is suitable for real-time deployment.

LanBLoc-2L: Two-Landmark Localization

Extending LanBLoc to operate with fewer landmarks by resolving geometric ambiguity through candidate generation and image-retrieval-based disambiguation.

Cross-View Geo-Localization

Exploring alignment between drone and satellite views for large-scale UAV geo-localization using viewpoint-invariant feature learning and coarse-to-fine retrieval.

Model Benchmarking

Systematic evaluation of detection architectures for landmark recognition:

• YOLO variants (v5–v12)
• Transformer-based detectors (RT-DETR)
• Two-stage models (Faster R-CNN)

Focused on precision, recall, mAP, inference latency, and deployment feasibility.

Research Vision

My long-term goal is to develop scalable, infrastructure-free localization systems that operate reliably in real-world, constrained, and adversarial environments – fully passive, hybrid learning + geometry, deployable on edge devices.